[
https://issues.apache.org/jira/browse/HBASE-2027?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Dave Latham updated HBASE-2027:
-------------------------------
Attachment: 2027-LRU.patch
I spent some more time thinking about this issue, and reading through the
background HBASE-1251 which set it up.
The cache of connections as it stands now is never reduced or freed due to the
strong reference each TableServers holds back to the HBaseConfiguration key in
the HBASE_INSTANCES map. However, with the first patch I provided, if client
code keeps a strong reference to an HBaseConfiguration but not to the
connection itself, then the connection information may be freed even if the
HBaseConfiguration object is still around. This is not desirable either.
Another possibility would be to convert the reference the TableServers holds to
the HBaseConfiguration to a WeakReference instead of converting the
HBASE_INSTANCES value to holding a WeakReference to the TableServers. However,
this also presents problems because then if the client held a strong reference
to the HConnection but not to the HBaseConfiguration (less likely than the
earlier case, I believe) then the configuration reference could be freed and
methods that require it would fail.
I propose another simpler method, to get rid of the WeakHashMap /
WeakReferences entirely and make the HBASE_INSTANCES map a simple LRU cache of
the last 10 HBaseConfiguration to HConnection instances. This will bound the
amount of memory used up by the cache (better than the current implementation)
at the risk of hbase clients who need to repeatedly use more than 10 different
hbase configurations losing cached connections (I believe this to be very rare,
but please enlighten me if I am mistaken.)
I'm attaching a new patch with this solution. Do you think it would be better
to make the size of this cache a config option instead of hardcoding it?
> HConnectionManager.HBASE_INSTANCES leaks TableServers
> -----------------------------------------------------
>
> Key: HBASE-2027
> URL: https://issues.apache.org/jira/browse/HBASE-2027
> Project: Hadoop HBase
> Issue Type: Bug
> Affects Versions: 0.20.0
> Reporter: Dave Latham
> Assignee: Dave Latham
> Fix For: 0.20.3, 0.21.0
>
> Attachments: 2027-LRU.patch, 2027.patch
>
>
> HConnectionManager.HBASE_INSTANCES is a WeakHashMap from HBaseConfiguration
> to TableServers. However, each TableServers has a strong reference back to
> the HBaseConfiguration key so they are never freed. (See note at
> http://java.sun.com/javase/6/docs/api/java/util/WeakHashMap.html :
> "Implementation note: The value objects in a WeakHashMap are held by ordinary
> strong references. Thus care should be taken to ensure that value objects do
> not strongly refer to their own keys, either directly or indirectly, since
> that will prevent the keys from being discarded.")
> Moreover, HBaseConfiguration implements hashCode() but not equals() so
> identical HBaseConfiguration objects each get their own TableServers object.
> We had a long running HBase client process that was creating new HTable()
> objects, each creating a new HBaseConfiguration() and thus a new TableServers
> object. It eventually went OOM, and gave a heap dump indicating 360 MB of
> data retained by HBASE_INSTANCES.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
